The healthy ageing gene expression signature for Alzheimer's disease diagnosis: a random sampling perspective

The first part of the experiments in Sood et al. [2015] builds their signature and assesses its ability to predict chronological age in different settings. This first part involves eight gene expression studies: six from muscle, one from brain and one from skin samples. The skin study was done on Illumina Human HT-12 V3 arrays and all the others on Affymetrix HGU133plus2 arrays. The first muscle dataset involves muscle samples from 15 young and 15 old healthy individuals and is only used to build the signature. The selection process retains probe-sets which are both differentially expressed between young and old samples as measured by limma [Ritchie et al., 2015] and predictive of chronological age in the context of a 5 nearest neighbor classifier, along with other selected probe-sets. The 150 probe-sets selected constitute the healthy ageing gene signature (HAGS) and they are then used in a 5 nearest neighbor classifier to predict the chronological age of samples in the other studies; the study used to select the signature is not used anymore in the rest of their experiments. Sood et al. [2015] use two different protocols to evaluate the prediction performance. For all except the skin data, they use external validation: the samples from one of the muscle studies (Campbell) are used as neighbors to predict the age of the tested samples in the four remaining muscle and the brain study. For the skin study, they use leave one out cross validation (LOOCV). They also use LOOCV on two of the muscle studies and the brain