Can Genetic Programming Perform Explainable Machine Learning for Bioinformatics?

Abstract

Although proven powerful in making predictions and finding patterns, machine learning algorithms often struggle to provide explanations and translational knowledge when applied to many problems, especially in biomedical sciences. This is often resulted by the highly complex structure employed by machine learning algorithms to represent and model the relationship of the predictors and the response. The prediction accuracy is increased at the cost of having a black-box model that is not amenable for interpretation. Genetic programming may provide a potential solution to explainable machine learning for bioinformatics where learned knowledge and patterns can be translated to clinical actions. In this study, we employed an LGP algorithm for a bioinformatics classification problem. We developed feature selection analysis methods and aimed at explaining which features are influential in the prediction, and whether such an influence is through individual effects or synergistic effects of combining with other features.