Unified imputation of missing data modalities and features in multi-omic data via shared representation learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Multi-omic studies offer comprehensive biological insights but suffer from incomplete data, with missing modalities and feature-level missingness.
• Existing imputation methods are limited, addressing either missing modalities or missing values, but not both simultaneously.

Purpose of the Study:

• To introduce MIMIR, a unified deep learning framework for reconstructing both missing data modalities and feature-level missing values in multi-omic datasets.
• To develop a method capable of handling arbitrary combinations of missing modalities and feature imputation.

Main Methods:

• MIMIR employs shared representation learning, utilizing modality-specific masked autoencoders to learn representations.
• These representations are projected into a common latent space, enabling data reconstruction from any observed subset of modalities.
• The framework was evaluated on The Cancer Genome Atlas (TCGA) pan-cancer multi-omic data.

Main Results:

• MIMIR consistently outperformed baseline methods in various missing-modality and missing-value scenarios, including missing completely at random (MCAR) and missing not at random (MNAR) settings.
• Analysis of the learned shared space revealed structured cross-modal dependencies, with transcriptional and epigenetic data forming a core, and copy number variation providing distinct signals.
• Imputation accuracy varied across modalities, influenced by learned cross-modal relationships.

Conclusions:

• Shared representation learning provides an effective and flexible foundation for unified multi-omic imputation under heterogeneous missingness.
• MIMIR successfully addresses the dual challenges of missing modalities and feature-level missingness, advancing the field of multi-omic data integration.
• The framework's ability to capture cross-modal dependencies enhances understanding of biological systems and imputation accuracy.