ACE: a versatile contrastive learning framework for single-cell mosaic integration

Abstract

The integration of single-cell multi-omics datasets is critical for deciphering cellular heterogeneities. Mosaic integration, the most general integration task, poses a greater challenge regarding disparity in modality abundance across datasets. Here, we present ACE, a mosaic integration framework that assembles two types of strategies to handle this problem: modality-alignment based strategy (ACE-align) and regression-based strategy (ACE-spec). ACE-align utilizes a novel contrastive learning objective for explicit modality alignment to uncover the shared latent representations behind modalities. ACE-spec combines the modality-alignment results and modality-specific representations to construct complete multi-omics representations for all datasets. Extensive experiments across various mosaic integration scenarios demonstrate the superiority of ACEs two strategies over existing methods. Application of ACE-spec to bi-modal and tri-modal integration scenarios showcases that ACE-spec is able to enhance the representation of cellular heterogeneities for datasets with incomplete modalities. The source code of ACE can be accessed at https://github.com/CSUBioGroup/ACE-main.