MetDIT

MetDIT integrates deep learning approaches to transform one-dimensional (1D) metabolomics sequence data into two-dimensional (2D) image representations and perform CNN-based feature extraction for improved classification and interpretation of clinical metabolomics datasets.

Key Features:

• TransOmics: Transforms one-dimensional (1D) sequence data into two-dimensional (2D) images while maintaining precise correspondence between original sequences and image representations.
• NetOmics: Employs a convolutional neural network (CNN) architecture to extract highly discriminative features from the transformed 2D images.
• Feature Augmentation Module (FAM): Implements feature augmentation to mitigate overfitting arising from small sample sizes and to address class imbalance.
• Optimized loss function: Incorporates an optimized loss function to improve model performance and robustness.
• Backbone and image resolution optimization: Systematically optimizes model backbone and image resolution to balance computational efficiency and parameter complexity.
• Benchmarking: Experimentally validated on three clinical metabolomics datasets with performance comparisons to Random Forest, SVM, XGBoost, and LightGBM.

Scientific Applications:

• Clinical metabolomics classification: Enables classification of clinical metabolomics datasets to support precision medicine analyses.
• High-dimensional data analysis: Addresses challenges of high-dimensional, intercorrelated metabolomic datasets that are difficult for traditional machine learning methods.
• Feature discovery: Facilitates identification of subtle patterns and relationships in metabolomics data through CNN-derived discriminative features.

Methodology:

MetDIT converts 1D sequence data to 2D images via TransOmics, applies NetOmics (CNN) for feature extraction, uses a feature augmentation module and an optimized loss function to mitigate overfitting and class imbalance, systematically optimizes model backbone and image resolution, and validates classification performance against Random Forest, SVM, XGBoost, and LightGBM on three clinical metabolomics datasets.