Liam tackles complex multimodal single-cell data integration challenges

Area of Science:

• Single-cell biology
• Computational biology
• Genomics

Background:

• Multi-omics characterization of single cells offers insights into gene regulatory dynamics.
• Integrating multimodal single-cell data, especially from diverse sources with technical variation, remains a challenge.

Purpose of the Study:

• To introduce liam, a flexible model for integrating paired single-cell multimodal data.
• To enable both horizontal and vertical integration, as well as mosaic integration with unimodal data.
• To learn a joint low-dimensional representation beneficial for data with varying information content or quality.

Main Methods:

• Developed liam, a model for simultaneous horizontal and vertical integration of paired single-cell multimodal data.
• Implemented mosaic integration of paired with unimodal data.
• Utilized a combination of conditional and adversarial training to account for complex batch effects, optimizing with replicate information.

Main Results:

• Demonstrated superior performance of liam on paired multimodal data types like Multiome and CITE-seq.
• Showcased liam's effectiveness in mosaic integration scenarios.
• Highlighted the model's ability to retain selected biological variation while correcting for technical noise.

Conclusions:

• liam provides a flexible and effective solution for integrating diverse single-cell multimodal datasets.
• The model's joint representation learning and batch effect correction advance the analysis of gene regulatory states.
• Benchmarking reveals remaining challenges and opportunities in single-cell data integration assessment.