Biomarker Modelling in Omics Technologies Using Symbolic Regression

Abstract

Omics data can contain predictive information of the onset of diseases and chronic conditions. Applying machine learning (ML) techniques to omics data is a promising venue of research, but domain data sets are typically high-dimensional and low-sample-size, presenting significant challenges to classic ML approaches. Another obstacle is the black-box nature of many ML algorithms, which prevents them from being deployed in medical practice. Symbolic regression (SR) is a possible solution to obtain human-interpretable models; but even equations cannot be easily understood, if they include hundreds or thousands of features. While feature selection can help reducing the number of features to be considered, most algorithms make unrealistic assumptions or bias the selection using a single classifier. In this work, we apply the Recursive Ensemble Feature Selection (REFS) algorithm, designed to avoid over-relying on a single ML model, with a modern SR algorithm, to obtain interpretable models predictive for different diseases, starting from real-world omics data. Experimental results for five different omics studies show that the completely open-source approach is competitive with the state-of-the-art in closed-source software. Comparing the same pipeline with REFS and more classic feature selection techniques shows that models created with REFS have a better performance.