Permutation-based Identification of Important Biomarkers for Complex Diseases via Black-box Models

Study of human disease remains challenging due to convoluted disease etiologies and complex molecular mechanisms at genetic, genomic, and proteomic levels. Many machine learning-based methods, including deep learning and random forest, have been developed and widely used to alleviate some analytic challenges in complex human disease studies. While enjoying the modeling flexibility and robustness, these model frameworks suffer from non-transparency and difficulty in interpreting the role of each individual feature due to their intrinsic black-box natures. However, identifying important biomarkers associated with complex human diseases is a critical pursuit towards assisting researchers to establish novel hypotheses regarding prevention, diagnosis and treatment of complex human diseases. Herein, we propose a Permutation-based Feature Importance Test (PermFIT) for estimating and testing the feature importance, and for assisting interpretation of individual feature in various black-box frameworks, including deep neural networks, random forests, and support vector machines. PermFIT (available at https://github.com/SkadiEye/deepTL) is implemented in a computationally efficient manner, without model refitting for each permuted data. We conduct extensive numerical studies under various scenarios, and show that PermFIT not only yields valid statistical inference, but also helps to improve the prediction accuracy of black-box models with top selected features. With the application to the Cancer Genome Atlas (TCGA) kidney tumor data and the HITChip atlas BMI data, PermFIT clearly demonstrates its practical usage in identifying important biomarkers and boosting performance of black-box predictive models.