scMMGAN

scMMGAN integrates multiple single-cell data modalities into a unified ambient data space to enable cross-modal alignment and analysis of complex omics datasets.

Key Features:

• Multi-modal integration: Integrates multiple single-cell data modalities into a cohesive representation within an ambient data space.
• Adversarial learning (GAN): Employs generative adversarial networks to learn cross-modal mappings via adversarial learning.
• Diffusion geometry loss: Incorporates a diffusion geometry loss to regularize mappings and improve geometric consistency.
• Novel kernel: Uses a novel kernel within the diffusion geometry loss to manage potential over-parameterization of GANs.
• Improved alignment precision: Produces more precise and meaningful alignments across diverse data modalities compared to existing methods.
• Spatial pattern discovery: Enables identification of spatial patterns within single-cell datasets.
• Demonstrated on disease data: Applied to triple-negative breast cancer single-cell datasets to reveal spatial patterns.

Scientific Applications:

• Spatial pattern analysis: Reveals spatial patterns in single-cell datasets, as shown in triple-negative breast cancer studies.
• Cross-modal downstream analysis: Provides a unified representation that facilitates comprehensive downstream analyses of omics data.
• Cellular and disease insights: Supports investigation of cellular behaviors and disease mechanisms, including cancer.

Methodology:

scMMGAN combines adversarial learning with generative adversarial networks and a diffusion geometry loss that employs a novel kernel to integrate multiple single-cell modalities into an ambient data space and produce aligned representations.