A penalized integrative deep neural network for variable selection among multiple omics datasets.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Deep learning models are increasingly used for omics data analysis, with variable selection enhancing interpretability.
• Existing deep learning methods struggle with small sample sizes common in omics data and may yield inaccurate results when pooling data from multiple studies due to structural differences.

Purpose of the Study:

• To develop a novel penalized integrative deep neural network (PIN) for simultaneous variable selection across multiple omics datasets.
• To address limitations of existing methods in handling small sample sizes and cross-dataset heterogeneity in integrative omics analysis.

Main Methods:

• Proposed a penalized integrative deep neural network (PIN) that aggregates multiple datasets as input.
• PIN accounts for both homogeneous and heterogeneous variable structures across different datasets within an integrative framework.
• Implemented simultaneous variable selection to identify important features across diverse omics data.

Main Results:

• Extensive simulations and real-world applications demonstrated PIN's superior performance compared to existing methods.
• PIN achieved considerably improved variable selection accuracy across multiple datasets, outperforming naive data pooling.
• The method was successfully applied to gene expression datasets from studies on cognitive status in elders and ovarian cancer staging.

Conclusions:

• The proposed PIN method effectively identifies important disease-related variables from multiple omics datasets.
• PIN offers a robust solution for integrative omics analysis, handling small sample sizes and cross-dataset variations.
• The freely available source code (rucliyang/PINFunc) facilitates the adoption of PIN in future multi-study omics research.