MOGEDN: small-sample cancer subtype classification with encoder-decoder networks for missing-omics recovery and...

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Cancer subtype classification using multi-omics data is challenging due to missing data and limited samples.
• Existing methods like graph convolutional networks (GCNs) struggle with entirely missing omics modalities.

Purpose of the Study:

• To develop a novel framework, MOGEDN, for robust cancer subtype classification from incomplete multi-omics data.
• To enhance classification performance in small-sample and missing-omics scenarios.

Main Methods:

• MOGEDN utilizes multi-omics encoder-decoder networks to reconstruct latent features of missing omics data.
• A step-wise algorithm pretrains the model on diverse cancer types and finetunes it for specific cancers, integrating inter-sample and cross-omics dependencies.
• The reconstructed and available omics features are integrated for prediction.

Main Results:

• MOGEDN consistently outperforms state-of-the-art baselines in accuracy and F1 scores on TCGA cancer datasets, especially for subtypes with fewer than 50 samples.
• The framework demonstrates robust performance in small-sample and missing-omics settings.
• Feature analysis yields two sets of complementary biomarkers: shared across cancer types and specific to a cancer type.

Conclusions:

• Decoder-based imputation is a powerful approach for multi-omics learning, improving classification accuracy and few-shot performance.
• MOGEDN facilitates multi-scale biomarker discovery, enhancing model interpretability and biological insights.
• The framework offers a robust solution for analyzing incomplete multi-omics data in cancer research.