Interpretable and integrative analysis of single-cell multiomics with scMKL

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Single-cell analysis methods face a trade-off between predictive accuracy and interpretability, especially for multimodal data.
• Complex machine learning models offer high accuracy but lack transparency, while simpler models are interpretable but less predictive.
• Integrating diverse single-cell omics data (e.g., scRNA-seq, ATAC-seq) is crucial for comprehensive biological insights.

Purpose of the Study:

• To introduce a novel method, single-cell Multiple Kernel Learning (scMKL), for multimodal single-cell data analysis.
• To merge the predictive power of complex models with the interpretability of linear methods for actionable insights.
• To improve the classification of healthy versus cancerous cell populations across various cancer types.

Main Methods:

• Developed and applied scMKL, a Multiple Kernel Learning framework, to integrate single-cell RNA sequencing, ATAC sequencing, and 10x Multiome data.
• Utilized scMKL for classifying cell populations in breast, lymphatic, prostate, and lung cancers.
• Employed cross-dataset learning to leverage insights from one dataset for analysis in another.

Main Results:

• scMKL demonstrated superior performance in classifying healthy and cancerous cell populations compared to existing methods across multiple cancer types.
• The method successfully identified key transcriptomic, epigenetic, and multimodal pathways distinguishing cancer subtypes and progression.
• Interpretable results pinpointed critical biological features that were previously unachieved by other methods.

Conclusions:

• scMKL offers a powerful and interpretable approach for multimodal single-cell data integration and analysis.
• The method enhances understanding of cancer biology by uncovering novel pathways related to treatment response, tumor grade, and subtype classification.
• scMKL provides actionable insights for cancer research and clinical applications by bridging the gap between predictive accuracy and biological interpretability.