Generalizable Single-cell Multimodal Data Integration with Self-supervised Learning.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics technologies allow simultaneous measurement of diverse cellular data.
• Integrating multimodal single-cell data presents challenges like overfitting in small datasets and poor generalization in large atlases.

Purpose of the Study:

• To develop a unified deep learning framework for robust single-cell multimodal data integration.
• To address limitations of existing methods in handling both small-scale and large-scale datasets.

Main Methods:

• Developed MINERVA (multimodal integration with self-supervised learning), a deep learning framework.
• Employed self-supervised learning strategies for single-cell multimodal integration.
• Evaluated performance against six state-of-the-art methods.

Main Results:

• MINERVA demonstrated superior performance in dimensionality reduction, missing feature imputation, and batch effect correction, even with limited cells.
• Constructed scalable multi-tissue references for large-scale applications.
• Achieved zero-shot knowledge transfer, instant cell type annotation, and novel cell state identification without retraining.

Conclusions:

• MINERVA effectively bridges small-scale precision with atlas-level generalization in single-cell research.
• Serves as a versatile tool for de novo data integration and cost-effective reuse of existing atlases.
• Facilitates comprehensive downstream analyses for single-cell multimodal data.