scACT: Accurate Cross-modality Translation via Cycle-consistent Training from Unpaired Single-cell Data

Area of Science:

• Genomics
• Computational Biology
• Single-cell Omics

Background:

• Single-cell sequencing enables multi-modal profiling, crucial for understanding cellular regulation.
• Cross-modality translation methods are limited by the need for scarce, high-quality co-assay data.

Purpose of the Study:

• To introduce scACT, a deep generative model for extracting cross-modality biological insights from unpaired single-cell data.
• To address challenges in aligning unpaired data, enabling translation without prior knowledge, and exploring regulatory networks.

Main Methods:

• scACT utilizes adversarial training for unpaired multi-modal data alignment.
• Cycle-consistent training facilitates cross-modality translation without prior knowledge.
• In-silico perturbations enable interpretable exploration of regulatory interconnections.

Main Results:

• scACT demonstrated superior performance across diverse single-cell datasets in all three core tasks.
• The model successfully aligned unpaired data and performed cross-modality translation effectively.
• Interpretable regulatory insights were gained through in-silico perturbation analyses.

Conclusions:

• scACT advances single-cell omics analysis by enabling cross-modality translation from unpaired data.
• The developed open-source software package facilitates broader research community adoption.
• This method overcomes data limitations, paving the way for deeper biological discoveries.