Multi-scale deep learning for the imbalanced multi-label protein subcellular localization prediction based on...

Area of Science:

• Biophysics
• Computational Biology
• Bioinformatics

Background:

• Microscopic imaging advances protein subcellular location studies.
• Existing image-based prediction methods face challenges with feature representation, imbalanced data, and multi-label classification.

Purpose of the Study:

• To propose MSTLoc, a novel multi-scale deep learning model for protein subcellular location prediction.
• To address limitations in current methods, particularly for imbalanced multi-label immunohistochemistry (IHC) image datasets.

Main Methods:

• MSTLoc employs a deep convolutional neural network for multi-scale feature extraction from IHC images.
• Feature fusion aggregates high-level and low-level features to capture subcellular location dependencies.
• Vision Transformer (ViT) models feature relationships to enhance representation ability.

Main Results:

• MSTLoc demonstrates superior performance compared to state-of-the-art models in multi-label subcellular location prediction.
• Learned multi-scale deep features effectively capture discriminative patterns, outperforming hand-crafted features.
• Feature visualization confirms complementarity of multi-scale features for performance improvement.
• Case studies successfully identified cancer-related protein biomarkers.

Conclusions:

• MSTLoc offers an effective solution for protein subcellular location prediction, especially on imbalanced multi-label IHC data.
• The model's ability to learn discriminative multi-scale features enhances predictive accuracy.
• The developed webserver provides convenient access to the MSTLoc tool.