Deep Mining from Omics Data.

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Omics data analysis

Background:

• High-throughput omics technologies generate vast molecular data (genes, proteins, metabolites).
• Deep neural networks (deep nets) are increasingly preferred for complex omics data modeling due to superior performance over traditional methods.
• A significant challenge is the inherent lack of transparency in deep nets, hindering clinical trust despite accurate predictions.

Purpose of the Study:

• To summarize deep neural network architectures used in omics research.
• To provide a comprehensive overview of deep feature mining techniques for interpreting deep nets.
• To enhance the transparency and trustworthiness of deep learning models in omics.

Main Methods:

• Categorization of deep feature mining techniques into three groups: hidden layer visualization, input feature importance, and output layer gradient analysis.
• Review of methods for interpreting hidden layer weights and node activations.
• Discussion of deconvolutional network-based approaches and bespoke attribute impact measures.

Main Results:

• Omics researchers have made progress in interpreting deep nets by analyzing hidden layers to identify salient input features.
• Existing methods offer insights into model behavior and feature relevance.
• Further approaches are needed to fully elucidate the relationships between input data, hidden representations, and output predictions.

Conclusions:

• Interpreting deep neural networks is crucial for building trust in their application to omics data.
• Various deep feature mining techniques aid in understanding model decision-making processes.
• Continued development of interpretation methods is essential for advancing the reliable use of deep learning in biological and clinical research.