Elastic Coupled Co-clustering for Single-Cell Genomic Data

Pengcheng Zeng, Zhixiang Lin

Abstract

The recent advances in single-cell technologies have enabled us to profile genomic features at unprecedented resolution and datasets from multiple domains are available, including datasets that profile different types of genomic features and datasets that profile the same type of genomic features across different species. These datasets typically have different powers in identifying the unknown cell types through clustering, and data integration can potentially lead to a better performance of clustering algorithms. In this work, we formulate the problem in an unsupervised transfer learning framework, which utilizes knowledge learned from auxiliary dataset to improve the clustering performance of target dataset. The degree of shared information among the target and auxiliary datasets can vary, and their distributions can also be different. To address these challenges, we propose an elastic coupled co-clustering based transfer learning algorithm, by elastically propagating clustering knowledge obtained from the auxiliary dataset to the target dataset. Implementation on single-cell genomic datasets shows that our algorithm greatly improves clustering performance over the traditional learning algorithms.

Introduction

Identifying distinct cell types through clustering is a cornerstone of single-cell analysis. However, researchers often face "difficult" datasets—those with high noise levels, small sample sizes, or poor feature representations.

While we often have access to "richer" auxiliary datasets (like a well-mapped human atlas or high-resolution scRNA-seq data), traditional clustering algorithms treat every dataset as an island. Existing integration methods often require the datasets to be very similar. The challenge addressed here is: How can we leverage knowledge from a high-quality auxiliary dataset to improve clustering in a target dataset, even when they come from different species, measure different genomic features, or have different underlying distributions?

Methods: The elasticC3 Framework

The authors propose Elastic Coupled Co-clustering (elasticC3). Unlike standard clustering which only groups cells, co-clustering groups both cells and features (like genes or genomic regions) simultaneously.

Conceptual workflow of the two-step elastic coupled co-clustering process.

The "Elastic" approach works in two main steps:

  1. Auxiliary Co-clustering: The algorithm first performs information-theoretic co-clustering on the auxiliary dataset to establish a "knowledge base" of cell and feature clusters.
  2. Target Co-clustering with Transfer: When clustering the target data, the algorithm adds "elastic" constraints (using Kullback-Leibler divergence) to its objective function. These constraints encourage the target clusters to resemble the auxiliary clusters.
Detailed schematic of elasticC3 showing the transfer of clustering knowledge from auxiliary to target data.

The "elasticity" is controlled by hyperparameters ($\alpha$ and $\beta$) that determine how much weight to give the auxiliary knowledge. This allows the model to adaptively decide how much information to transfer based on how related the two datasets actually are. Crucially, the model remains functional even if the number of cell types differs between the two datasets.

Datasets and Validation

The researchers validated their model using both synthetic and real-world biological data:

  • Simulated Data: Modeled to mimic scRNA-seq and scATAC-seq data with varying levels of correlation between features.
  • Real Data 1 (Cross-Modality): Human scRNA-seq (gene expression) used as auxiliary data to improve the clustering of Human scATAC-seq (chromatin accessibility) for K562 and HL60 cell lines.
  • Real Data 2 (Cross-Species): Human scRNA-seq used as auxiliary data to improve the clustering of Mouse scRNA-seq for lung cells (alveolar type 2, clara, and ependymal cells).

Key Insights

  • Co-clustering is a Force Multiplier: By clustering features alongside cells, the model captures the "why" behind cell groupings, making the transfer of knowledge across domains (like Human to Mouse) much more effective.
  • Elasticity Prevents "Negative Transfer": If an auxiliary dataset is unrelated, a rigid transfer would hurt performance. The elastic nature of the algorithm allows it to "loosen" the influence of the auxiliary data if it doesn't fit the target data's distribution.
  • Handling "Missing" Information: The model is robust. In tests where a specific cell type existed in the target data but was removed from the auxiliary data, elasticC3 still outperformed traditional methods, showing it doesn't over-rely on a perfect 1:1 match between datasets.

Results

The elasticC3 algorithm consistently outperformed state-of-the-art clustering methods (like SC3 and SIMLR) and traditional transfer learning approaches (like STC).

  • Accuracy Boost: In the human-to-mouse lung cell experiment, elasticC3 achieved an Adjusted Rand Index (ARI) of 0.682, significantly higher than the 0.434 achieved by the next best transfer method (STC) and the 0.625 of standard co-clustering.
  • Efficiency: The algorithm is computationally efficient, with the authors demonstrating empirical convergence within just 10 iterations across various datasets.

Conclusion and Future Work

The authors acknowledge that finding the global optimal solution for co-clustering is NP-hard, meaning the algorithm converges to a local minimum. While their results are strong, the final outcome can depend on the initial starting point.

Additionally, the model currently requires a grid search to tune the hyperparameters ($\alpha, \beta$, and the number of feature clusters $K$). Future work could involve automating this tuning process. The authors also suggest that the "elastic" framework is general enough to be applied beyond genomics, potentially finding use in fields like text mining.

The source code and data sets are available at https://github.com/cuhklinlab/elasticC3.

Summary updated

Abstract

The recent advances in single-cell technologies have enabled us to profile genomic features at unprecedented resolution and datasets from multiple domains are available, including datasets that profile different types of genomic features and datasets that profile the same type of genomic features across different species. These datasets typically have different powers in identifying the unknown cell types through clustering, and data integration can potentially lead to a better performance of clustering algorithms. In this work, we formulate the problem in an unsupervised transfer learning framework, which utilizes knowledge learned from auxiliary dataset to improve the clustering performance of target dataset. The degree of shared information among the target and auxiliary datasets can vary, and their distributions can also be different. To address these challenges, we propose an elastic coupled co-clustering based transfer learning algorithm, by elastically propagating clustering knowledge obtained from the auxiliary dataset to the target dataset. Implementation on single-cell genomic datasets shows that our algorithm greatly improves clustering performance over the traditional learning algorithms. The source code and data sets are available at https://github.com/cuhklinlab/elasticC3.
BibTeX
@misc{zeng2020elastic,
  author = {Pengcheng Zeng and Zhixiang Lin},
  title = {{Elastic Coupled Co-clustering for Single-Cell Genomic Data}},
  year = {2020},
  eprint = {2003.12970},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.12970},
}
Figure