Dynamic Synthesis of Multi‐Modal Representations for CITE‐seq Data Integration and Analysis

The accurate integration of CITE‐seq modalities faces challenges due to complex RNA‐protein nonlinear interactions and computational inefficiencies. A lightweight framework that adaptively fuses RNA and ADT modalities, and refines them through fine‐grained learning is developed to address this issue. This model enhances multi‐modal integration and clustering performance, revealing dynamics of immune cell development in mouse thymocytes. Abstract Single‐cell multi‐omics technologies are pivotal for deciphering the complexities of biological systems, with Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE‐seq) emerging as a particularly valuable approach. The dual‐modality capability makes CITE‐seq particularly advantageous for dissecting cellular heterogeneity and understanding the dynamic interplay between transcriptomic and proteomic landscapes. However, existing computational models for integrating these two modalities often struggle to capture the complex, non‐linear interactions between RNA and antibody‐derived tags (ADTs), and are computationally intensive. To address these issues, scMHVA, a novel and lightweight framework designed to integrate the diverse modalities of CITE‐seq data, is proposed. scMHVA utilizes an adaptive dynamic synthesis module to generate consolidated yet heterogeneous embeddings from RNA and ADT modalities. Subsequently, scMHVA enhances inter‐modality correlations within the joint representation by applying a multi‐head self‐attention mechanism, effectively capturing the intricate mapping relationships between mRNA expression levels and protein abundance. Extensive experiments demonstrate that scMHVA consistently outperformed existing single‐modal and multi‐modal clustering methods across CITE‐seq datasets of varying scales, exhibiting linear runtime scalability and effectively eliminating batch effects, thereby establishing it as a robust tool for large‐scale CITE‐seq data analysis. Additionally, it is demonstrated that scMHVA successfully annotates different cell types in a published mouse thymocyte dataset and reveals dynamics of immune cell development. The accurate integration of CITE-seq modalities faces challenges due to complex RNA-protein nonlinear interactions and computational inefficiencies. A lightweight framework that adaptively fuses RNA and ADT modalities, and refines them through fine-grained learning is developed to address this issue. This model enhances multi-modal integration and clustering performance, revealing dynamics of immune cell development in mouse thymocytes. Abstract Single-cell multi-omics technologies are pivotal for deciphering the complexities of biological systems, with Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE-seq) emerging as a particularly valuable approach. The dual-modality capability makes CITE-seq particularly advantageous for dissecting cellular heterogeneity and understanding the dynamic interplay between transcriptomic and proteomic landscapes. However, existing computational models for integrating these two modalities often struggle to capture the complex, non-linear interactions between RNA and antibody-derived tags (ADTs), and are computationally intensive. To address these issues, scMHVA, a novel and lightweight framework designed to integrate the diverse modalities of CITE-seq data, is proposed. scMHVA utilizes an adaptive dynamic synthesis module to generate consolidated yet heterogeneous embeddings from RNA and ADT modalities. Subsequently, scMHVA enhances inter-modality correlations within the joint representation by applying a multi-head self-attention mechanism, effectively capturing the intricate mapping relationships between mRNA expression levels and protein abundance. Extensive experiments demonstrate that scMHVA consistently outperformed existing single-modal and multi-modal clustering methods across CITE-seq datasets of varying scales, exhibiting linear runtime scalability and effectively eliminating batch effects, thereby establishing it as a robust tool for large-scale CITE-seq data analysis. Additionally, it is demonstrated that scMHVA successfully annotates different cell types in a published mouse thymocyte dataset and reveals dynamics of immune cell development. Advanced Science, Volume 12, Issue 44, November 27, 2025.