MIC_Locator

MIC_Locator predicts protein subcellular localization from image data using multi-label classification and frequency-domain feature extraction to support proteomics and genome annotation.

Key Features:

• Multi-Label Prediction Model: A model explicitly developed to handle multi-label localization datasets rather than single-label predictions.
• Updated Benchmark Dataset: Trained and evaluated on a contemporary benchmark dataset containing updated label information.
• Advanced Feature Extraction: Extracts frequency features from the three components of monogenic signals at multiple scales using Fourier transformation, Riesz transformation, and Log-Gabor filters.
• Intensity Encoding Strategy: Applies an intensity coding/encoding strategy to complement frequency-domain features and capture localization detail.
• Chained Prediction Model: Employs a chained prediction approach to improve handling of multi-label relationships.
• Performance: Reports a subset accuracy of 60.56% and outperforms many existing models by leveraging frequency features over traditional spatial-domain or grey-level descriptors.

Scientific Applications:

• Protein Function Validation: Supports validation of protein function by providing predicted subcellular localization information.
• Genome Annotation: Provides localization annotations to assist genome annotation efforts and identification of molecular targets.
• Drug Design and Development: Informs drug design by identifying protein locations within specific subcellular compartments relevant to target selection.

Methodology:

Frequency feature extraction focused on frequency-domain descriptors rather than spatial or grey-level descriptors; monogenic signal decomposition into three components at multiple scales using Fourier transformation, Riesz transformation, and Log-Gabor filters; an intensity coding strategy; and a chained multi-label prediction model trained on an updated benchmark dataset.