Recover then aggregate: unified cross-modal deep clustering with global structural information for single-cell data

Ziyi Wang, Peng Luo, Mingming Xiao, Boyang Wang, Tianyu Liu, Xiangyu SunBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces UCMDC, a novel deep clustering framework for single-cell multi-modal data that tackles the pervasive issue of missing values and inconsistent cross-modal representations. By first recovering missing data and then aggregating information using an attention-based mechanism and contrastive learning, UCMDC achieves superior clustering performance compared to existing state-of-the-art methods.

What Problem Does This Paper Address?

Single-cell sequencing technologies generate high-resolution data crucial for understanding cellular heterogeneity. However, analyzing this data, especially when combining information from multiple sources (multi-modal data), presents significant challenges. A primary issue is the presence of numerous missing values (dropouts) in gene expression matrices, which, if not handled properly, can lead to inaccurate clustering. Many existing methods treat these missing values as zeros or perform imputation as a separate preprocessing step, failing to integrate this recovery effectively into the clustering process. Furthermore, current cross-modal clustering approaches often struggle with inconsistent representations across different data modalities, potentially leading to conflicting information and suboptimal results. This paper aims to address these limitations by proposing a unified framework that simultaneously recovers missing data and integrates information from multiple modalities in a consistent and effective manner.

Methods

The proposed UCMDC model employs a "Recover then Aggregate" strategy.

  • Discriminative Data Recovery: This module addresses missing values by assessing the confidence of zero values. It uses Non-negative Matrix Factorization (NMF) guided by a recovery confidence metric and a guidance matrix to selectively impute missing data. This approach aims to avoid introducing noise from over-recovery.
  • Cross-View Aggregation: This module uses an attention mechanism, inspired by transformers, to aggregate representations from different modalities. It constructs a global structural relationship matrix to assign weights to each modality, allowing for more efficient data integration.
  • Contrastive Learning Optimization: This module aligns modality-specific representations with consistent ones using contrastive loss. It pulls samples belonging to the same cluster closer together while simultaneously optimizing with a reconstruction loss to generate high-quality cell representations.

The entire framework is jointly optimized for both reconstruction and clustering losses.

Datasets

The study evaluated UCMDC on four publicly available single-cell multi-modal datasets:

  • Stuart
  • Pbmc
  • Gayoso
  • Smage

These datasets vary in the number of cells, modalities, categories, and feature dimensions.

Key Insights and Ideas

  • Unified Recovery and Aggregation: The core innovation is integrating data recovery directly into the deep clustering framework, rather than treating it as a separate preprocessing step. This allows the recovery process to be guided by the downstream clustering objective.
  • Confidence-Based Recovery: Instead of blindly imputing all missing values, UCMDC intelligently assesses the confidence of zero values, leading to more robust data recovery and preventing the introduction of spurious signals.
  • Structure-Guided Aggregation: The use of an attention mechanism and a global structural relationship matrix allows for a more nuanced aggregation of information from different modalities, accounting for their varying quality and importance.
  • Contrastive Learning for Alignment: Applying contrastive learning helps to align representations across modalities, ensuring that cells within the same cluster are represented consistently, regardless of their origin modality.

Key Results

  • Superior Clustering Performance: UCMDC consistently outperformed eight state-of-the-art multi-modal clustering methods across multiple evaluation metrics (ACC, NMI, ARI) on all four datasets. For instance, UCMDC achieved a 94.16% ACC on the Stuart dataset, significantly higher than the next best method.
  • Effective Data Recovery: Ablation studies demonstrated that the data recovery module significantly improved clustering performance, with improvements ranging from 0.29% to 35.78% across various metrics and datasets when compared to UCMDC without the recovery module.
  • Module Contributions: Ablation studies also showed that both the cross-modal aggregation and contrastive learning modules were crucial for UCMDC's performance, with their removal leading to substantial declines in clustering accuracy.

Limitations and Future Work

  • Uniform Recovery Confidence: The current recovery confidence algorithm is applied uniformly across all modalities, without considering their distinct distributions. This could lead to suboptimal data recovery in some cases.
  • Future Directions: The authors plan to develop customized recovery confidence algorithms for different modalities and explore more efficient methods for cross-modal information aggregation to further refine the model. They also aim to address the growing complexity of biological data analysis.

Summary updated

BibTeX
@article{wang2024recover,
  author = {Ziyi Wang and Peng Luo and Mingming Xiao and Boyang Wang and Tianyu Liu and Xiangyu Sun},
  title = {{Recover then aggregate: unified cross-modal deep clustering with global structural information for single-cell data}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae485},
}
Figure