Imbalanced classification for protein subcellular localization with multilabel oversampling.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in proteomics

Background:

• Accurate human protein subcellular localization is crucial for understanding cellular functions and disease mechanisms.
• Computational prediction of protein localization faces challenges due to data imbalance, where minority classes are underrepresented, leading to biased models.
• Multilabel classification settings exacerbate data imbalance issues due to the co-occurrence of majority and minority classes.

Purpose of the Study:

• To develop and evaluate a novel oversampling method to improve the prediction accuracy of human protein subcellular localization, particularly for minority classes.
• To address the complex data imbalance problem in multilabel classification of protein localization.

Main Methods:

• Quantified data imbalance and class concurrence in the Human Protein Atlas dataset.
• Developed a new oversampling strategy incorporating non-linear mixup, geometric, and color transformations for data augmentation.
• Implemented a specialized sampling approach for minibatch preparation.

Main Results:

• The proposed oversampling method significantly boosts classification performance for minority protein classes.
• Demonstrated superior predictions for minority classes compared to existing methods on the Human Protein Atlas dataset.
• Validated the effectiveness of data augmentation and tailored sampling in handling imbalanced multilabel data.

Conclusions:

• Oversampling minority samples via advanced data augmentation techniques is a promising strategy for improving protein localization prediction.
• The developed method offers a robust solution for handling data imbalance in multilabel classification tasks.
• This work contributes to more accurate computational tools for protein function and disease research.