Permutation-based identification of important biomarkers for complex diseases via machine learning models

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Human disease research faces challenges due to complex molecular mechanisms and genetic factors.
• Machine learning models offer powerful analytical tools but often lack transparency, hindering biomarker identification.
• Interpreting individual features in complex models is crucial for advancing disease understanding and therapeutic strategies.

Purpose of the Study:

• To develop and validate a computationally efficient method, Permutation-based Feature Importance Test (PermFIT), for estimating and testing feature importance in complex machine learning frameworks.
• To enhance the interpretability of sophisticated algorithms like deep neural networks, random forests, and support vector machines.
• To aid researchers in identifying critical biomarkers for complex human diseases.

Main Methods:

• Developed Permutation-based Feature Importance Test (PermFIT) for feature importance estimation and testing.
• Implemented PermFIT in a computationally efficient manner, avoiding model refitting.
• Applied PermFIT to various machine learning models, including deep neural networks, random forests, and support vector machines.

Main Results:

• PermFIT provides valid statistical inference for feature importance.
• The method improves the prediction accuracy of machine learning models.
• Extensive numerical studies demonstrated PermFIT's effectiveness across diverse scenarios.

Conclusions:

• PermFIT offers a practical and efficient approach to interpreting complex machine learning models in biomedical research.
• The tool successfully identified important biomarkers in real-world datasets, such as the Cancer Genome Atlas kidney tumor data and the HITChip atlas data.
• PermFIT aids in hypothesis generation for disease prevention, diagnosis, and treatment by enhancing biomarker discovery and model performance.