Systematic auditing is essential to debiasing machine learning in biology.

Area of Science:

• Life Sciences
• Computational Biology
• Biomedical Applications

Background:

• Biases in training data are common in biological datasets.
• These biases can inflate machine learning (ML) model performance and obscure learning processes.
• Systematic auditing of ML models for bias is not standard practice in life sciences.

Purpose of the Study:

• To develop a systematic and generalizable framework for auditing ML models in life sciences.
• To identify and address unrecognized biases affecting ML model performance in therapeutic applications.
• To understand the impact of data biases on ML model learning when biological signals are weak.

Main Methods:

• Devised a systematic, principled, and general auditing framework for ML models.
• Applied the framework to examine three ML applications with therapeutic relevance.
• Analyzed model performance and learning patterns in the presence of data biases.

Main Results:

• Identified unrecognized biases in ML models that significantly hinder the learning process.
• Demonstrated that biases can lead to substantially reduced model performance on new datasets.
• Showed that ML models predominantly learn from data biases when insufficient biological signal is present.

Conclusions:

• A systematic auditing approach is crucial for reliable ML application in life sciences.
• Addressing data biases is essential to improve ML model accuracy and interpretability.
• The developed framework and protocols can be adapted for various biomedical ML applications.