Matching single cells across modalities with contrastive learning and optimal transport

Area of Science:

• Computational Biology
• Genomics
• Biotechnology

Background:

• Understanding biomolecular interactions is crucial for cellular behavior.
• Single-cell technologies enable multimodal data acquisition, but present challenges like high dimensionality and noise.
• Computational methods for cross-modality cell matching are needed to interpret complex single-cell data.

Purpose of the Study:

• To develop a novel computational method, MatchCLOT, for accurate and efficient modality matching in multimodal single-cell data.
• To leverage contrastive learning and optimal transport for learning common representations and performing cell matching.
• To address the challenges of high dimensionality and noise in single-cell datasets.

Main Methods:

• MatchCLOT utilizes contrastive learning to learn shared representations between different data modalities.
• Entropic optimal transport is employed as an approximate maximum weight bipartite matching algorithm.
• The method was evaluated on curated benchmarking and independent test datasets.

Main Results:

• MatchCLOT achieved state-of-the-art performance, outperforming existing methods by 26.1% on benchmark datasets.
• The method effectively preserves the underlying biological structure of multimodal single-cell data.
• MatchCLOT demonstrates significant improvements in computational time and memory efficiency, enabling scalability for large datasets.

Conclusions:

• MatchCLOT provides an accurate and efficient solution for modality matching in multimodal single-cell analysis.
• The method's scalability makes it suitable for increasingly large single-cell datasets.
• MatchCLOT advances the interpretation of cellular complexity and heterogeneity from multimodal single-cell data.