Self-omics: A Self-supervised Learning Framework for Multi-omics Cancer Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Next Generation Sequencing provides vast multi-omics data, but high dimensionality and lack of annotation pose analysis challenges.
• Self-Supervised Learning (SSL) addresses limited labeled data in machine learning, yet its application to inter-omics relationships in unlabeled data is underexplored.

Purpose of the Study:

• To develop a novel and efficient SSL pre-training paradigm for multi-omics data analysis.
• To leverage inter-omics relationships within unlabeled data to enhance downstream machine learning tasks.
• To improve cancer type classification and feature extraction from high-dimensional omics datasets.

Main Methods:

• Developed a pre-training paradigm incorporating contrastive alignment, data recovery from corrupted samples, and cross-omic data imputation.
• Utilized various Self-Supervised Learning (SSL) components for effective representation learning.
• Designed network architectures to handle missing omic types and multiple datasets.

Main Results:

• The proposed SSL pre-training paradigm significantly improves performance on downstream tasks with limited labeled data.
• Outperformed state-of-the-art methods in semi-supervised cancer type classification on the TCGA pancancer dataset.
• Pre-trained encoders demonstrated effectiveness as feature extractors without fine-tuning.
• Ablation studies confirmed the robustness and non-dependency on specific pretext tasks.

Conclusions:

• The novel SSL pre-training paradigm effectively addresses challenges in multi-omics data analysis.
• The method offers a powerful approach for cancer type classification and feature extraction, even with limited labeled data.
• The paradigm shows potential for extension to zero-shot classification of rare cancers and robust handling of missing data.