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

Area of Science:

• Single-cell multi-omics
• Computational biology
• Immunology

Background:

• Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE-seq) enables simultaneous measurement of RNA and protein levels.
• Existing methods struggle to capture complex interactions between RNA and antibody-derived tags (ADTs) and are computationally intensive.
• Accurate integration of multi-modal CITE-seq data is crucial for understanding cellular heterogeneity.

Purpose of the Study:

• To develop a novel, lightweight computational framework, scMHVA, for integrating diverse CITE-seq data modalities.
• To capture complex, non-linear interactions between transcriptomic and proteomic data.
• To provide a robust and scalable tool for large-scale CITE-seq data analysis.

Main Methods:

• scMHVA employs an adaptive dynamic synthesis module to create consolidated embeddings from RNA and ADT data.
• A multi-head self-attention mechanism is utilized to enhance inter-modality correlations and capture mRNA-protein relationships.
• The framework was evaluated on CITE-seq datasets of varying scales.

Main Results:

• scMHVA outperformed existing single-modal and multi-modal clustering methods.
• The framework demonstrated linear runtime scalability and effective batch effect removal.
• scMHVA successfully annotated cell types in a mouse thymocyte dataset and revealed immune cell development dynamics.

Conclusions:

• scMHVA is an effective and efficient tool for integrating CITE-seq data, advancing the analysis of cellular heterogeneity.
• The framework's ability to capture transcriptomic-proteomic interplay offers new insights into biological systems.
• scMHVA provides a scalable solution for analyzing large CITE-seq datasets in immunology and beyond.