AggMapNet: enhanced and explainable low-sample omics deep learning with feature-aggregated multi-channel networks.

Area of Science:

• Biomedical informatics
• Computational biology
• Machine learning

Background:

• Omics data presents high-dimensional and low-sample size challenges for deep learning.
• Efficient algorithms are crucial for analyzing limited omics datasets in biomedical research.

Purpose of the Study:

• To develop a novel unsupervised feature aggregation tool (AggMap) for omics data.
• To create advanced deep learning models (AggMapNet) for enhanced omics data analysis and interpretability.

Main Methods:

• AggMap aggregates omics features into multi-channel 2D image-like feature maps (Fmaps) based on intrinsic correlations.
• AggMapNet, a multi-channel deep learning model, utilizes AggMap Fmaps as input.
• The Simply-explainer module provides interpretability for AggMapNet predictions.

Main Results:

• AggMap demonstrated superior feature reconstruction capabilities on benchmark datasets.
• AggMapNet models outperformed state-of-the-art machine learning methods on 18 low-sample omics tasks.
• AggMapNet showed enhanced robustness with noisy data and improved disease classification accuracy, identifying key biomarkers for COVID-19.

Conclusions:

• The AggMap-AggMapNet pipeline offers a powerful approach for enhanced learning and interpretability of low-sample omics data.
• This method addresses key challenges in omics-based biomedical research, particularly with limited sample sizes.
• The integration of unsupervised feature engineering and supervised explainable deep learning provides a robust framework for future omics studies.