multiVIB: A unified probabilistic contrastive learning framework for atlas-scale integration of single-cell...

Area of Science:

• Neuroscience
• Computational Biology
• Genomics

Background:

• Comprehensive brain cell atlases are crucial for understanding neural functions and translational medicine.
• The proliferation of single-cell technologies generates diverse datasets requiring robust integration methods.
• Existing data integration tools are often limited, necessitating complex, artifact-prone workflows.

Purpose of the Study:

• To introduce multiVIB, a unified probabilistic contrastive learning framework for integrating heterogeneous single-cell brain data.
• To address the need for scalable and biologically faithful data integration across diverse experimental platforms, species, and modalities.
• To overcome limitations of existing tools and prevent spurious data alignments.

Main Methods:

• Development of multiVIB, a probabilistic contrastive learning framework.
• Application of multiVIB to atlas-scale datasets from the BRAIN Initiative.
• Evaluation of multiVIB's performance in integrating diverse data modalities and cross-species datasets.

Main Results:

• multiVIB achieves state-of-the-art performance in data integration.
• The framework effectively mitigates spurious alignments between datasets.
• Demonstrated robust and scalable integration of diverse data modalities and preservation of species-specific variations in cross-species integration.

Conclusions:

• multiVIB provides a unified, scalable, and biologically faithful framework for constructing next-generation brain cell atlases.
• The framework supports the integration of heterogeneous single-cell data, crucial for advancing neuroscience.
• multiVIB enables reliable construction of brain atlases from the growing volume of single-cell data.