Enabling Large Language Models to Comprehend the Language of Single-cell Paired Multi-omics Data

Large language models (LLMs) have recently shown great potential for single-cell analysis, offering powerful performance, natural-language interpretability, and flexible usability. However, existing LLM-based single-cell methods remain largely centered on single-omics (single-modal) data, and their extension to paired multi-omics (multi-modal) data remains challenging. An empirical study reveals a counterintuitive ”more modalities, worse performance” phenomenon: directly providing paired multi-omics data to LLMs fails to improve performance and can even lead to worse performance than single-omics inputs. We argue that this limitation stems from the prevailing text-level concatenation paradigm in this field, which serializes heterogeneous multi-omics data into discrete text and simply concatenates them as inputs to an LLM. This paradigm has fundamental limitations and cannot effectively integrate complex and heterogeneous multi-omics information. To address this challenge, we propose **LLM M**ulti-**O**mics **Reader** (**LLM-MOReader**), a framework that shifts LLM-based paired single-cell multi-omics analysis from the text-level concatenation paradigm to explicit multi-omics fusion and projection paradigm. LLM-MOReader first uses variational inference to jointly model paired multi-omics data and learn a fused multi-omics representation, which is then mapped into LLM token space through continuous multi-omics prompt, enabling LLM-based reasoning over already-integrated multi-omics information. Experiments on diverse datasets across multiple species, modality combinations, and downstream tasks, including cell type annotation and multi-omics integration, show that LLM-MOReader matches or outperforms existing methods, including domain-specific LLMs, frontier large-scale LLMs and specialized non-LLM models. Further analyses show that our approach resolves the “more modalities, worse performance” problem and exhibits model-agnostic generality. LLM-MOReader enables LLMs to read the language of multi-omics, making LLM-based paired single-cell multi-omics analysis practically feasible.