Vec2image: an explainable artificial intelligence model for the feature representation and classification of...

Area of Science:

• Bioinformatics
• Artificial Intelligence
• Genomics

Background:

• Machine learning on large biological datasets faces challenges in model interpretability.
• Effective feature engineering and selection are crucial for biological data analysis.

Purpose of the Study:

• To develop an explainable artificial intelligence (AI) framework for biological data analysis.
• To improve model determination and prediction with mechanistic explanations.
• To enhance feature selection and classification for biological datasets.

Main Methods:

• Developed Vec2image, an explainable convolutional neural network framework.
• Utilized principal component coordinate conversion, deep residual neural networks, and k-nearest neighbor representation.
• Created pseudo images from high-dimensional biological data to represent features and associations.

Main Results:

• Vec2image demonstrated superior performance compared to existing methods.
• Successfully applied to cell marker identification in single-cell datasets.
• Achieved robust classification for type 2 diabetes (T2D) using human islet scRNA-seq data.
• Identified T2D-relevant genes, uncovering potential cellular pathogenesis and dysfunctions.

Conclusions:

• Vec2image is an efficient and explainable AI methodology for biological data.
• Enables human-readable classification and prediction using pseudo image representations.
• Offers insights into disease mechanisms by analyzing cell states and gene relevance.