LSMMD-MA

LSMMD-MA performs scalable multimodal alignment of single-cell omics datasets by reformulating MMD-MA (Maximum Mean Discrepancy for Multimodal Alignment) to enable modality matching across datasets containing up to a million cells per modality.

Key Features:

• Scalable MMD-MA reformulation: Implements a scalable version of MMD-MA (Maximum Mean Discrepancy for Multimodal Alignment) to enable large-scale multimodal integration.
• Large-scale capability: Handles datasets up to approximately one million cells per modality, providing orders-of-magnitude improvement over previous implementations.
• Optimization and computation: Reformulates the optimization using linear algebra and leverages KeOps, a CUDA-based framework for symbolic matrix computation in Python, for efficient processing of massive datasets.

Scientific Applications:

• Multimodal single-cell integration: Aligns cells across genomic assays (single-cell omics) to integrate transcriptomic, epigenomic, and other modalities at single-cell resolution.
• Cross-assay modality matching for biology and clinical research: Enables unification of insights from different genomic technologies to study complex biological systems and support precision diagnostics and treatments.

Methodology:

Reformulates the MMD-MA optimization using linear algebra and solves it with KeOps, a CUDA-based symbolic matrix computation framework for Python.