TextOmics: Language-Driven Omics Integration

• TextOmics is a paradigm that treats natural language as a central component in omics analysis, enabling text-guided retrieval and integration of biochemical data.
• It employs multi-scale alignment techniques—such as optimal transport and graph-based embedding—to connect free-text with molecules, signaling graphs, and genomic outputs.
• Empirical evaluations demonstrate strong gains in cross-modal retrieval, cell annotation, and generative tasks, while also highlighting challenges in scalability and interpretability.

TextOmics is a research paradigm in which natural language is treated as a first-class component of omics analysis, molecular representation, or biological reasoning. Across recent work, the term covers several related formulations: using free text as a query interface to molecules and biochemical substructures; embedding human-readable annotations inside signaling graphs built from single-cell data; automating omics pipelines and literature reuse with LLM agents; translating omics profiles into marker-gene text or structured biological descriptions; and jointly conditioning molecular generation on omics expressions and textual constraints (Min et al., 2024, Zhang et al., 2 Apr 2025, Mondal et al., 2024, Liu et al., 2024, Li et al., 5 Feb 2025, Yuan et al., 14 Jul 2025, Kharouiche et al., 29 Sep 2025, Hutton et al., 10 Mar 2026, Sypetkowski et al., 7 May 2026).

1. Conceptual scope

TextOmics does not denote a single architecture or benchmark. In cross-modal cheminformatics, it places natural language at the interface to molecules, metabolites, fragments, and higher-order biochemical structures, with the goal of precise text–molecule alignment and retrieval. In single-cell systems biology, it denotes the integration of human-readable biomedical knowledge—names, descriptions, functions, locations, diseases, drugs, sequences, and chemistry—with quantitative omics features inside signaling graphs. In workflow automation, it refers to systems in which an LLM both executes an omics pipeline and produces source-cited interpretation. In executable-literature settings, it extends to agentic systems that transform publications, supplements, and repository links into searchable research objects with provenance, downloadable data, and rerunnable analyses (Min et al., 2024, Zhang et al., 2 Apr 2025, Mondal et al., 2024, Hutton et al., 10 Mar 2026).

A second major branch of TextOmics treats language as either the output space or the reasoning substrate for omics. SOAR converts single-cell profiles into textual marker lists and evaluates instruction-tuned LLMs for zero-shot cell type annotation. Cell2Text maps scRNA-seq profiles to structured natural-language descriptions of cell identity, tissue, disease, and pathway activity. OmicsLM instead inserts compact continuous omics representations directly into an LLM context so that multiple biological samples can be reasoned over jointly with textual instructions. Omni-DNA extends the same general idea to genomic foundation models by enabling a single decoder to map DNA sequences to labels, free-form text, or discretized images (Liu et al., 2024, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026, Li et al., 5 Feb 2025).

The name also appears as a specific heterogeneous benchmark for molecular generation. In that formulation, TextOmics establishes one-to-one correspondences between omics expressions, molecular textual descriptions, and SELFIES representations, creating triple-aligned samples of the form $$(\mathbf{E} \in \mathbb{R}^{978}, \mathbf{C}, \text{SELFIES})$$ for controllable hit-like molecular generation (Yuan et al., 14 Jul 2025).

2. Representation strategies and architectural patterns

A recurring TextOmics pattern is explicit multi-granular alignment between language and structured biochemical representations. ORMA uses a SciBERT text encoder and a hierarchical heterogeneous molecular graph encoder with atom, motif, and molecule nodes. Token embeddings $z_i = h_i^t$ and motif embeddings $m_j = h_j^m$ are coupled through an optimal transport objective with cosine-based cost

$C_{ij} = 1 - \cos(z_i, m_j),$

and ORMA approximately solves

$\min_{T \in U(a,b)} \langle T, C \rangle$

with IPOT. The learned transport plan is converted into motif-aligned multi-token phrase representations

$h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$

after which contrastive objectives are applied at token–atom, multi-token–motif, and sentence–molecule scales. The total loss is

$$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$

with $\alpha = 0.5$ and $\beta = 0.2$ in training. This is presented as the first attempt to explore alignments at both the motif and multi-token levels (Min et al., 2024).

Another strategy embeds text directly into graphs rather than aligning it post hoc. OmniCellTOSG formalizes a text-omic signaling graph as

$G = (V, E, X, T, A, Y),$

where nodes are transcripts and proteins, edges represent internal signaling or PPIs, numeric node features encode expression or abundance, and text annotations are derived from BioMedGraphica. Text encoders such as DeBERTa, DNAGPT, and ProtGPT2 generate node-level text features, which are fused with numeric features through

$z_i = h_i^t$0

A representative message-passing update is

$z_i = h_i^t$1

In this formulation, text is not merely metadata; it is part of the graph state propagated during GNN computation (Zhang et al., 2 Apr 2025).

A different architectural family translates omics into an LLM-readable interface. SOAR serializes each cell type as a compact text list of the top $z_i = h_i^t$2 DEGs, ranked by Welch’s $z_i = h_i^t$3-test, and prompts instruction-tuned LLMs with zero-shot or two-stage chain-of-thought templates. Cell2Text retains per-gene resolution by passing Geneformer embeddings $z_i = h_i^t$4 through a two-layer adapter

$z_i = h_i^t$5

and then optimizes the LLM with

$z_i = h_i^t$6

OmicsLM rejects long gene-name serialization and instead inserts one continuous token per sample. Its composite omics vector is

$z_i = h_i^t$7

which is projected into the Qwen3 token space via

$z_i = h_i^t$8

This produces a single multimodal context in which several biological samples and textual instructions can be interleaved (Liu et al., 2024, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026).

Omni-DNA and ToDi show that TextOmics also encompasses generative decoders. Omni-DNA pretrains a decoder-only transformer on DNA sequences with next-token prediction,

$z_i = h_i^t$9

then expands its vocabulary with task prompts, label tokens, natural-language text tokens, and VQ-VAE image tokens. ToDi combines a VAE-based omics encoder, a frozen SciBERT text encoder, and a diffusion generator over remapped SELFIES. Its diffusion backbone fuses semantic and omics conditions as

$m_j = h_j^m$0

and trains with a joint reconstruction, token-level NLL, and cosine alignment objective (Li et al., 5 Feb 2025, Yuan et al., 14 Jul 2025).

3. Task families and application domains

One major TextOmics task is cross-modal retrieval. Text-to-molecule retrieval returns chemical structures from free-text descriptions, while the reverse direction retrieves descriptions from molecules. ORMA formulates both directions as cross-modal ranking over large candidate sets, motivated by chemical database search, metabolite annotation, and literature–structure linking. The motivating examples are explicitly phrase-level, such as “pyridyl ring” or “hydroxymethyl side chain,” rather than only sentence-level semantics (Min et al., 2024).

A second task family concerns cell and tissue interpretation. SOAR treats cell type annotation as a text-grounded inference problem in which marker genes are mapped to ontology-normalized cell types, and extends this to scATAC-seq through cross-modality translation into RNA-like profiles. OmniCellTOSG supports cell type classification, condition or status prediction, organ or tissue classification, and disease classification over signaling graphs whose nodes carry both numeric and textual attributes. Cell2Text moves beyond discrete labels by generating structured descriptions that summarize cellular identity, tissue of origin, disease context, and pathway activity, while OmicsLM expands the scope to multi-sample biological question answering over real transcriptomic studies (Liu et al., 2024, Zhang et al., 2 Apr 2025, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026).

A third family uses text as the orchestrator of omics workflows. SeqMate is designed for one-click analytics starting from raw FASTQ files, automating trimming with cutadapt, QC graphics with bio, alignment with hisat, SAM-to-BAM conversion with pysam, quantification with featureCounts, differential expression with pydeseq2, and then generation of a cited report from PubMed, PDB, UniProt, NCBI GenBank, and Ensembl. Omics Data Discovery Agents applies a comparable principle at the publication level: agents identify articles, extract metadata, locate raw or processed datasets, configure and execute containerized pipelines through MCP, and publish research objects that record metadata, checksums, parameters, logs, and downstream analysis outputs (Mondal et al., 2024, Hutton et al., 10 Mar 2026).

TextOmics is also used for sequence-to-language and omics-conditioned molecular generation. Omni-DNA introduces DNA2Function, in which DNA sequences are mapped to textual functional descriptions, and Needle-in-DNA, in which sequences are mapped to images through a shared decoder. ToDi targets hit-like molecular generation by conditioning on both transcriptomic context and molecular descriptions, and it extends this setup to target perturbation and zero-shot therapeutic generation from disease signatures plus symptom narratives. In this branch of the literature, text is both a constraint language for generation and a semantic layer that shapes chemical validity, substructure composition, and biological plausibility (Li et al., 5 Feb 2025, Yuan et al., 14 Jul 2025).

4. Benchmarks and empirical findings

On text–molecule retrieval, ORMA reports strong gains on both ChEBI-20 and PCdes. On ChEBI-20, text→molecule retrieval reaches Hits@1 $m_j = h_j^m$1, Hits@10 $m_j = h_j^m$2, MRR $m_j = h_j^m$3, and Mean Rank $m_j = h_j^m$4; molecule→text retrieval reaches Hits@1 $m_j = h_j^m$5, Hits@10 $m_j = h_j^m$6, MRR $m_j = h_j^m$7, and Mean Rank $m_j = h_j^m$8. On PCdes, ORMA reports text→molecule $m_j = h_j^m$9, $C_{ij} = 1 - \cos(z_i, m_j),$0, $C_{ij} = 1 - \cos(z_i, m_j),$1, and MRR $C_{ij} = 1 - \cos(z_i, m_j),$2, with molecule→text $C_{ij} = 1 - \cos(z_i, m_j),$3, $C_{ij} = 1 - \cos(z_i, m_j),$4, $C_{ij} = 1 - \cos(z_i, m_j),$5, and MRR $C_{ij} = 1 - \cos(z_i, m_j),$6. Its ChEBI-20 ablations also show that combining token–atom, multi-token–motif, and sentence–molecule losses is superior to removing any one scale, while MM-only training yields text→molecule Hits@1 $C_{ij} = 1 - \cos(z_i, m_j),$7 and Mean Rank $C_{ij} = 1 - \cos(z_i, m_j),$8 (Min et al., 2024).

In generative chemistry and genomic generation, the reported gains are similarly large. On ChemInduced, full ToDi reports Validity $C_{ij} = 1 - \cos(z_i, m_j),$9, Uniqueness $\min_{T \in U(a,b)} \langle T, C \rangle$0, Novelty $\min_{T \in U(a,b)} \langle T, C \rangle$1, Levenshtein $\min_{T \in U(a,b)} \langle T, C \rangle$2, FCD $\min_{T \in U(a,b)} \langle T, C \rangle$3, Morgan $\min_{T \in U(a,b)} \langle T, C \rangle$4, and MACCS $\min_{T \in U(a,b)} \langle T, C \rangle$5; its text-only and omics-only ablations are weaker. Omni-DNA achieves state-of-the-art performance on 18 out of 26 tasks across the Nucleotide Transformer and Genomic Benchmark suites. The paper reports an average of $\min_{T \in U(a,b)} \langle T, C \rangle$6 for Omni-DNA $\min_{T \in U(a,b)} \langle T, C \rangle$7 on the 18 NT tasks and $\min_{T \in U(a,b)} \langle T, C \rangle$8 for Omni-DNA $\min_{T \in U(a,b)} \langle T, C \rangle$9 on GB. On DNA2Function, Omni-DNA@ft reaches weighted F1 $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$0 and MCC $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$1. On Needle-in-DNA, Omni-DNA $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$2 achieves macro F1 $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$3 with $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$4 invalid generations on average (Yuan et al., 14 Jul 2025, Li et al., 5 Feb 2025).

Single-cell and multimodal language benchmarks show that TextOmics systems are not confined to retrieval or generation. SOAR evaluates eight instruction-tuned LLMs across 11 datasets and finds that explicit chain-of-thought can materially improve open models: Mixtral-8×22B rises from BLEU average $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$5 to $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$6, and Qwen2-72B increases BLEU-1 from $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$7 to $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$8. For scATAC-to-RNA translation, SOAR reports mean $h_j^p = \frac{1}{|D_j|}\sum_{i \in D_j} h_i^t,$9 across marker genes. Cell2Text, trained on 1,000,000 cells from 7,331 donors, reports cell type accuracy $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$0 for Gemma-4B and $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$1 for Llama-1B full fine-tuning, compared with $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$2 for Geneformer+Head; tissue accuracy reaches $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$3, disease accuracy $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$4, and average PageRank similarity reaches $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$5. OmicsLM introduces GEO-OmicsQA, a 3,000-example publication-disjoint benchmark grounded in real GEO transcriptomes, and reports aggregate Binary QA $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$6 and Free-text $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$7, together with GTEx validation performance of Purity $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$8 and Tissue $$L_{\text{total}} = \alpha L_{ta} + \beta L_{mm} + (1-\alpha-\beta)L_{sm},$$9, and zero-shot perturbation Recall@50 $\alpha = 0.5$0 on X-Atlas/Orion (Liu et al., 2024, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026).

Graph-grounded and agentic systems emphasize different metrics. OmniCellTOSG aggregates 117,519,978 raw single cells into 547,168 meta-cells and reports, for example, AD CellTOSG-Class accuracy of $\alpha = 0.5$1 for cell type and $\alpha = 0.5$2 for status, and SCLC accuracy of $\alpha = 0.5$3 for cell type and $\alpha = 0.5$4 for status. Omics Data Discovery Agents reports a benchmark of 39 proteomics publications in which 91 datasets were identified; excluding ambiguous cases, precision is $\alpha = 0.5$5 and recall $\alpha = 0.5$6, while for standard repositories 49 of 61 were correctly identified, described as approximately $\alpha = 0.5$7 precision. In a reanalysis of Taneera et al., overlap in differentially expressed proteins rises from $\alpha = 0.5$8 initially to $\alpha = 0.5$9 after matching software version and preprocessing choices. SeqMate, by contrast, reports no runtime, accuracy, or user-study benchmarks; its contribution is framed primarily as pipeline automation and cited interpretation rather than quantitative evaluation (Zhang et al., 2 Apr 2025, Hutton et al., 10 Mar 2026, Mondal et al., 2024).

5. Interpretability, evaluation, and recurrent misunderstandings

Interpretability is a central theme in TextOmics, but it is instantiated in different ways. ORMA makes phrase-level grounding explicit by transporting tokens to motifs and visualizing aligned substructures and phrases in matched colors; its case studies include (R)-nephthenol and L-histidinol. OmniCellTOSG locates interpretation in graph structure, attention-derived core signaling subgraphs, and the separation between internal signaling edges and PPI edges. Cell2Text exposes its predictions as structured narratives rather than only labels, and Omics Data Discovery Agents requires evidence spans for extracted metadata, records parameter manifests and checksums, and preserves execution provenance in research objects (Min et al., 2024, Zhang et al., 2 Apr 2025, Kharouiche et al., 29 Sep 2025, Hutton et al., 10 Mar 2026).

A common misunderstanding is to equate TextOmics with simple gene-name prompting. Some systems do serialize biology into text, and SOAR explicitly shows that compact DEG lists plus instruction-tuned LLMs can be effective. However, the broader literature repeatedly argues that text-only serialization can discard quantitative structure or mechanistic context. OmicsLM is explicit that ranked-gene strings sacrifice quantitative signal and saturate context windows, which is why it uses continuous projected omics tokens. OmniCellTOSG instead grounds text in signaling graphs, and ToDi uses one-to-one omics–text–SELFIES alignment rather than unpaired prompting. This suggests that TextOmics is better understood as a multimodal alignment program than as a purely textual reformulation of omics (Liu et al., 2024, Sypetkowski et al., 7 May 2026, Zhang et al., 2 Apr 2025, Yuan et al., 14 Jul 2025).

Another recurrent concern is whether free-form outputs weaken evaluation rigor. The published systems answer this by using task-specific metrics rather than relying on anecdotal generations alone: Hits@1, Hits@10, MRR, and Mean Rank for retrieval; BLEU, Exact Match, F1, and ontology-normalized evaluation for cell annotation; PageRank-based similarity on the Cell Ontology for generated cell descriptions; GEval-based judging for omics-grounded free-text QA; and FCD, Tanimoto, QED, novelty, and validity for molecule generation. In other words, TextOmics expands output spaces, but it does not necessarily abandon closed-form evaluation (Min et al., 2024, Liu et al., 2024, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026, Yuan et al., 14 Jul 2025).

6. Limitations, risks, and future directions

The limitations reported across TextOmics are heterogeneous but structurally related. ORMA notes that very long texts or molecules with many motifs increase OT cost, that BRICS decomposition can under- or over-segment chemotypes, and that the design currently targets small molecules rather than proteins, pathways, or multi-omics networks. OmniCellTOSG reports uneven organ and disease distributions, dependence on BioMedGraphica curation, and a current focus on transcriptomics plus virtual proteomic nodes. ToDi reports no wet-lab validation, only an Alzheimer’s disease case study for zero-shot therapeutic generation, and possible bias from LINCS cell lines, landmark-gene restriction, and functional-group imbalance (Min et al., 2024, Zhang et al., 2 Apr 2025, Yuan et al., 14 Jul 2025).

LLM-specific risks are also explicit. SeqMate warns about hallucinations, the absence of deterministic reproducibility controls, and privacy issues created by reliance on OpenAI’s gpt-3.5-turbo-0125 API. SOAR shows that explicit chain-of-thought can hurt very strong models such as GPT-4o by pushing them toward broader labels, and it identifies ncRNA-dominated marker sets and ontology mismatches as important failure modes. Cell2Text notes the possibility of hallucinated narratives, incomplete regex or synonym coverage for rare terms, and the lack of explicit batch-effect modeling in the architecture. OmicsLM states that it has no dedicated interpretability head, inherits biases from public datasets and GEO curation, and is not presented as clinically validated (Mondal et al., 2024, Liu et al., 2024, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026).

Agentic literature systems introduce additional security and provenance issues. Omics Data Discovery Agents is limited by PMC Open Access coverage, irregular supplement organization, incomplete or outdated parameter reporting, and prompt-injection risk. Its safeguards include separating text-ingesting agents from tool-executing agents, requiring abstention on low-confidence items, validating accession existence, and recording container digests, hashes, and logs. This suggests that, in TextOmics settings where text can trigger computation, provenance and tool isolation are not peripheral implementation details but core methodological requirements (Hutton et al., 10 Mar 2026).

The stated future directions are correspondingly broad. ORMA proposes unbalanced OT, Gromov–Wasserstein OT, and extension to proteins, pathways, and multi-omics graphs. OmniCellTOSG proposes richer modalities such as proteomics, spatial transcriptomics, ligand–receptor communication edges, and expanded literature-derived annotations. SOAR explicitly points to retrieval-augmented generation with resources such as CellMarker, PanglaoDB, UniProt, and the Cell Ontology. Omni-DNA emphasizes broader multimodal outputs and improved grounding with structured knowledge. Cell2Text points toward broader ontology and pathway coverage, while OmicsLM suggests richer conditioning and extension beyond human protein-coding panels. Taken together, these trajectories indicate that TextOmics is moving from prompt-level interpretation toward fully grounded multimodal systems that combine quantitative omics, structured biological knowledge, executable tools, and generative language in a single computational framework (Min et al., 2024, Zhang et al., 2 Apr 2025, Liu et al., 2024, Li et al., 5 Feb 2025, Kharouiche et al., 29 Sep 2025, Sypetkowski et al., 7 May 2026).