Extracting rules from trained machine learning models with applications in Bioinformatics

Recently, Machine learning methods have achieved great success in various areas. However, some machine learning-based models are not explainable (e.g., Artificial Neural Networks), which may affect the massive applications in medical fields.

In this talk, we first introduce two approaches that extract rules from trained neural networks. The first one leads to an algorithm that extracts rules in the form of Boolean functions. The second one extracts probabilistic rules representing relations between inputs and the output. We demonstrate the effectiveness of these two approaches by computational experiments.

Then we consider applying an explainable machine learning model to predict human Dicer cleavage sites. Human Dicer is an enzyme that cleaves pre-miRNAs into miRNAs. We develop an accurate and explainable predictor for the human Dicer cleavage site -- ReCGBM. Computational experiments show that ReCGBM achieves the best performance compared with several existing methods. Further, we find that features close to the center of pre-miRNA are more important for the prediction.