BOSO: A novel feature selection algorithm for linear regression with high-dimensional data

With the frenetic growth of high-dimensional datasets in different biomedical domains, there is an urgent need to develop predictive methods able to deal with this complexity. Feature selection is a relevant strategy in machine learning to address this challenge. We introduce a novel feature selection algorithm for linear regression called BOSO (Bilevel Optimization Selector Operator). We conducted a benchmark of BOSO with key algorithms in the literature, finding a superior accuracy for feature selection in high-dimensional datasets. Proof-of-concept of BOSO for predicting drug sensitivity in cancer is presented. A detailed analysis is carried out for methotrexate, a well-studied drug targeting cancer metabolism.Author summary: We present BOSO (Bilevel Optimization Selector Operator), a novel method to conduct feature selection in linear regression models. In machine learning, feature selection consists of identifying the subset of input variables (features) that are correctly associated with the response variable that is aimed to be predicted. An adequate feature selection is particularly relevant for high-dimensional datasets, commonly encountered in biomedical research questions that rely on -omics data, e.g. predictive models of drug sensitivity, resistance or toxicity, construction of gene regulatory networks, biomarker selection or association studies. The need of feature selection is emphasized in many of these complex problems, since the number of features is greater than the number of samples, which makes it harder to obtain accurate and general predictive models. In this context, we show that the models derived by BOSO make a better combination of accuracy and simplicity than competing approaches in the literature. The relevance of BOSO is illustrated in the prediction of drug sensitivity of cancer cell lines, using RNA-seq data and drug screenings from GDSC (Genomics of Drug Sensitivity in Cancer) database. BOSO obtains linear regression models with a similar level of accuracy but involving a substantially lower number of features, which simplifies the interpretation and validation of predictive models.