PScL-SDNNMAE: Protein Subcellular Localization Prediction Using Classical and Masked Autoencoder-Based Multi-View...

Area of Science:

• Computational Biology
• Cell Biology
• Bioinformatics

Background:

• Accurate protein subcellular localization is vital for cellular function and drug design.
• Existing computational methods lack sufficient performance in feature extraction.
• There is a need for efficient vision learners utilizing self-supervised learning for deep feature extraction.

Purpose of the Study:

• To propose PScL-SDNNMAE, a novel bioimage-based method for predicting protein subcellular localization in human cells.
• To enhance the accuracy and generalization capability of subcellular localization prediction.
• To leverage self-supervised learning for effective feature representation from bioimages.

Main Methods:

• Feature extraction using traditional image descriptors and a masked autoencoder (MAE) for deep features.
• Feature selection using Analysis of Variance (ANOVA), Mutual Information (MI), and stepwise discriminant analysis (SDA).
• Classification using a deep neural network (DNN) trained on an integrated feature set.

Main Results:

• PScL-SDNNMAE demonstrated superior performance and generalization compared to state-of-the-art predictors on benchmark experiments.
• 10-fold cross-validation and independent testing confirmed the method's effectiveness.
• Self-supervised learning proved effective for learning representations from IHC images.

Conclusions:

• PScL-SDNNMAE offers an advanced approach for protein subcellular localization prediction.
• The study highlights the potential of self-supervised learning in bioimage analysis.
• Future work may involve pre-training on large unlabeled datasets for further improvements.