Guiding questions to avoid data leakage in biological machine learning applications.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Biology

Background:

• Machine learning (ML) is crucial for pattern extraction in high-dimensional biological data.
• Reported ML prediction performance in biology often fails in real-world applications.
• Data leakage, or information sharing between training and test sets, is a primary cause of inflated performance estimates.

Purpose of the Study:

• To address the challenge of data leakage in biological machine learning.
• To provide a practical framework for preventing data leakage in biological datasets.
• To promote robust and reproducible machine learning research in the life sciences.

Main Methods:

• Development of seven critical questions to guide the prevention of data leakage.
• Application of these questions to non-trivial biological dataset examples.
• Illustrative analysis to demonstrate the utility of the proposed questions.

Main Results:

• Data leakage is a significant and often undetected issue in biological ML.
• The proposed seven questions effectively highlight potential data leakage scenarios.
• The framework aids in identifying and mitigating information contamination in biological models.

Conclusions:

• Awareness of potential data leakage is essential for biological ML.
• Implementing the seven questions can improve the reliability of ML models in biology.
• This approach supports the development of more robust and reproducible biological research using ML.