Title: Co-data weighted elastic net
Authors: Mark van de Wiel - Amsterdam University Medical Centers (Netherlands) [presenting]
Mirrelijn van Nee - Amsterdam University Medical Centers (Netherlands)
Abstract: Co-data stands for co-mplementary data. As opposed to covariates it contains information on the variables rather than on the samples. E.g. presence of a gene in a particular pathway, or a p-value that quantifies association strength between gene and outcome in a related, external study. Small sample size is a common challenge for clinical high-dimensional studies, e.g. due to cost restrictions or rarity of the disease. We show that the use of co-data can alleviate this challenge to some extent when the aim is to develop predictive signatures. From a methodological perspective, the focus lies on empirical Bayes techniques to incorporate the co-data information in the predictor, e.g. by adapting multiple penalty weights in the elastic net setting. Efficient computation of those penalties is a computational hurdle, which we take by an approximation that is of fairly general use. Ideas and software are illustrated by cancer genomics examples.