DULoc

DULoc estimates quantitative fractions of protein subcellular localization from immunofluorescence images using a deep-learning-based pattern unmixing approach.

Key Features:

• Deep-learning-based pattern unmixing: Employs a deep-learning-based pattern unmixing pipeline tailored for protein subcellular localization.
• Immunofluorescence input: Uses immunofluorescence images as the primary input modality for localization analysis.
• Deep convolutional neural network (CNN): Utilizes a deep convolutional neural network to construct detailed feature representations from images.
• Multiple nonlinear decomposing algorithms: Applies multiple nonlinear decomposing algorithms to process CNN-derived features during pattern unmixing.
• Quantitative fraction estimation: Quantitatively estimates fractions of proteins across multiple subcellular compartments for multi-label localization analysis.
• Validation on synthetic and real data: Validated on real and synthetic datasets, achieving correlation >0.93 between estimated and true protein fractions.
• Large-scale application to Human Protein Atlas: Applied to Human Protein Atlas data and matched existing annotations for 70.52% of proteins.

Scientific Applications:

• Quantitative subcellular localization: Estimation of protein localization fractions across subcellular compartments from immunofluorescence images.
• Multi-label protein analysis: Characterization of spatial distributions and functional mechanisms of multi-label proteins via fraction estimation.
• Proteome-scale annotation benchmarking: Benchmarking and alignment of localization predictions with Human Protein Atlas annotations.
• Image-based pattern unmixing: Unmixing overlapping subcellular patterns in immunofluorescence microscopy images.

Methodology:

Processes immunofluorescence images with a deep convolutional neural network to extract feature representations, then applies multiple nonlinear decomposing algorithms as part of a pattern unmixing pipeline to estimate protein localization fractions.