Explainable multi-task learning for multi-modality biological data analysis.

Area of Science:

• Single-cell multi-omics biology
• Computational biology
• Genomics and transcriptomics

Background:

• Biotechnology enables simultaneous measurement of multiple cellular modalities (RNA, DNA accessibility, protein).
• Analyzing multi-modal single-cell data requires integrating diverse analytical tasks for comprehensive understanding.
• Current methods are task-specific, offering incomplete insights into complex biological systems.

Purpose of the Study:

• To introduce UnitedNet, an explainable multi-task deep neural network for integrated analysis of single-cell multi-modality data.
• To demonstrate UnitedNet's capability in multi-modal integration and cross-modal prediction.
• To enable quantification of cell-type-specific relationships between gene expression and other modalities.

Main Methods:

• Development of UnitedNet, a multi-task deep neural network architecture.
• Application of UnitedNet to diverse single-cell multi-modality datasets (Patch-seq, multiome ATAC+gene expression, spatial transcriptomics).
• Utilizing explainable machine learning algorithms to dissect UnitedNet's predictions.

Main Results:

• UnitedNet achieved comparable or superior accuracy in multi-modal integration and cross-modal prediction against state-of-the-art methods.
• The framework successfully quantified cell-type-specific relationships between gene expression and other modalities.
• Demonstrated broad applicability across various single-cell multi-omics datasets.

Conclusions:

• UnitedNet provides a comprehensive end-to-end framework for analyzing single-cell multi-modality data.
• The explainable nature of UnitedNet facilitates direct quantification of regulatory relationships.
• UnitedNet has the potential to accelerate discoveries in cell-type-specific gene regulation across multiple modalities.