Abstract
Recent technological advances have led to an increase in the production and availability of single-cell data. The ability to integrate a set of multi-technology measurements would allow the identification of biologically or clinically meaningful observations through the unification of the perspectives afforded by each technology. In most cases, however, profiling technologies consume the used cells and thus pairwise correspondences between datasets are lost. Due to the sheer size single-cell datasets can acquire, scalable algorithms that are able to universally match single-cell measurements carried out in one cell to its corresponding sibling in another technology are needed. We propose Single-Cell data Integration via Matching (SCIM), a scalable approach to recover such correspondences in two or more technologies. SCIM assumes that cells share a common (low-dimensional) underlying structure and that the underlying cell distribution is approximately constant across technologies. It constructs a technology-invariant latent space using an auto-encoder framework with an adversarial objective. Multi-modal datasets are integrated by pairing cells across technologies using a bipartite matching scheme that operates on the low-dimensional latent representations. We evaluate SCIM on a simulated cellular branching process and show that the cell-to-cell matches derived by SCIM reflect the same pseudotime on the simulated dataset. Moreover, we apply our method to two real-world scenarios, a melanoma tumor sample and a human bone marrow sample, where we pair cells from a scRNA dataset to their sibling cells in a CyTOF dataset achieving 93% and 84% cell-matching accuracy for each one of the samples respectively.
Introduction
The "single-cell revolution" has provided unprecedented tools to examine cellular inner workings, from transcriptomics (gene expression) to proteomics (protein levels). However, a significant technical hurdle remains: most high-throughput profiling technologies are destructive. Consequently, researchers cannot perform multiple types of screens on the exact same physical cell. This results in "unpaired" datasets—different views of the same tissue sample with no direct record of which cell in one dataset corresponds to which cell in another.
Integrating these datasets is essential for understanding complex biological processes. Existing methods often struggle to scale to hundreds of thousands of cells or fail to capture complex, branching biological structures, such as stem cells differentiating into specialized types. SCIM (Single-Cell data Integration via Matching) addresses these challenges by providing a scalable framework to reconstruct a multi-modal view of biological samples.
Methodology
SCIM approaches the integration problem in two primary stages:
1. Constructing a Technology-Invariant Latent Space
SCIM utilizes an Adversarial Autoencoder (AAE) framework. Each technology (e.g., scRNA-seq or CyTOF) is assigned its own encoder/decoder pair. To ensure the latent representations are truly integrated, a discriminator network attempts to identify the source technology of a given latent code. The encoders are trained to "fool" this discriminator, effectively forcing disparate data types into a shared, overlapping coordinate system. This relies on the "sibling assumption": if two aliquots are taken from the same cell suspension, the underlying distribution of cell states should remain approximately constant.
2. Bipartite Matching
Once cells are mapped to the shared latent space, SCIM identifies "biological siblings" using a bipartite matching scheme. This is formulated as a Minimum-Cost Maximum-Flow problem. To ensure computational feasibility for massive datasets, SCIM employs a k-Nearest Neighbor (kNN) approximation to prune the search space, reducing complexity from $O(N^3)$ to a manageable level.
A critical innovation is the introduction of a "null node" ($Ø$) in the matching graph. This allows the algorithm to leave a cell unmatched if no biologically plausible partner is found, preventing the "forced matches" that often introduce noise in other integration tools.
Results and Validation
The authors validated SCIM across three distinct scenarios:
- Simulated Branching Process: Using 64,000 cells to mimic a 5-branch differentiation process, SCIM achieved a Pearson correlation of 0.86 in pseudotime values between matched cells, proving it preserves temporal progression.
- Melanoma Tumor Sample: Integrating scRNA-seq and CyTOF data, SCIM achieved 93% cell-matching accuracy. It successfully matched RNA expression to protein abundance for key markers like CD20 and CD3.
- Human Bone Marrow: Testing the ability to distinguish similar T-cell subpopulations, SCIM reached 84% accuracy.
Compared to state-of-the-art tools like MATCHER, which often assume simple one-dimensional trajectories, SCIM successfully handles high-dimensional, branching hierarchical structures common in real biology. Furthermore, SCIM demonstrates superior scalability, processing over 130,000 cells in roughly two hours.
Discussion and Future Work
SCIM represents a major advancement in multi-modal single-cell integration. While highly effective, the authors note that a small amount of supervision (as few as 10% of cell-type labels) is currently beneficial to prevent latent space orientation issues (e.g., mirroring). Future work will focus on making the method fully unsupervised for complex datasets and expanding its application to other modalities like spatial transcriptomics and single-cell ATAC-seq.
Availability
The SCIM framework is open-source and available at github.com/ratschlab/scim.
Summary updated