Interpreting how machine learning models make predictions in biological studies

ABSTRACT

Interpreting how the machine learning models make decisions is a new method to explore meaningful rules. However, it still lacks an understanding of the applicability of different model explainers in biological study. To address this question, we made a comprehensive evaluation on various explainers, and analyzed their performance and biological preference by quantifying the contribution of individual gene in the models trained to predict tissue type from transcriptome. Additionally, we also proposed a series of optimization strategies to improve the performance of different explainers. Interestingly, all explainers can be classified into three groups based on their outputs on different neural network architectures. With explainers from the group II, we found that the top contributing genes in different tissues exhibit tissue-specific manifestation and are potential biomarkers for cancer research. In summary, this work provides a novel insight and general guidance for exploring biological mechanisms by interpreting machine learning models.