CAPTAIN: a multimodal foundation model pretrained on co-assayed single-cell RNA and protein

Area of Science:

• Single-cell multi-omics analysis
• Computational biology
• Biomedical data science

Background:

• Proteins are key to cellular function, but transcriptomic data alone provides incomplete cell state information.
• Existing single-cell models trained only on transcriptomes yield biased and partial cellular characterizations.
• Integrating proteomic data is crucial for a comprehensive understanding of cellular phenotypes.

Purpose of the Study:

• To develop a multimodal foundational model for integrating single-cell transcriptomic and proteomic data.
• To overcome limitations of transcriptomic-only models in characterizing cellular states.
• To enable robust multi-omics integration and generate novel biological hypotheses.

Main Methods:

• Pretraining a multimodal foundational model (CAPTAIN) on over four million single cells with matched transcriptomic and surface protein data.
• Modeling cross-modality dependencies to learn unified representations.
• Evaluating CAPTAIN's performance in fine-tuning and zero-shot settings across diverse tissues.

Main Results:

• CAPTAIN learns unified multimodal representations, capturing cellular diversity across complex biological contexts.
• The model demonstrates robust generalization for protein imputation, cell type annotation, and batch harmonization.
• CAPTAIN successfully integrates multi-omics data, improving accuracy and generating new hypotheses on intercellular dynamics.

Conclusions:

• CAPTAIN offers a powerful approach for accurate single-cell multi-omics integration.
• The model enhances understanding of cellular states and protein-driven intercellular interactions.
• CAPTAIN has potential applications in disease research, such as identifying immune patterns in COVID-19 severity.