Toward Robust Self-Training Paradigm for Molecular Prediction Tasks.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Molecular prediction tasks often face challenges due to limited labeled data.
• Self-training is a semisupervised learning method that uses both labeled and unlabeled data, but teacher model-generated pseudo-labels can be inaccurate.
• Existing methods struggle with noisy pseudo-labels in self-training for molecular prediction.

Purpose of the Study:

• To develop a robust self-training strategy to address the issue of noisy pseudo-labels in molecular prediction.
• To enhance the performance of molecular prediction tasks by effectively utilizing both labeled and unlabeled data.
• To propose a universally applicable approach for the bioinformatics community.

Main Methods:

• Proposed a robust self-training strategy incorporating a robust loss function to handle noisy pseudo-labels.
• Explored two paradigms: generic and adaptive robust self-training.
• Evaluated the strategy on three molecular biology prediction tasks using four different backbone models.

Main Results:

• The robust self-training strategy significantly improved prediction performance across all tested molecular biology tasks.
• Molecular regression tasks showed a notable average enhancement of 41.5% in prediction performance.
• Visualization analysis confirmed the superiority of the proposed method over existing approaches.

Conclusions:

• The proposed robust self-training strategy is a simple yet effective method for improving molecular biology prediction.
• This approach successfully tackles the challenge of insufficient labeled data by utilizing unlabeled data.
• The method is easily embeddable in various prediction tasks, offering a universal solution for bioinformatics.