Background: Bioinformatics data, such as Genomics and metabolomics data are usually high dimensional and small sample size.This brings difficulty for researchers to explain the data and understand the complex biological process.Therefore many multivariate techniques and machine learning methods have been applied to analyze the bioinformatics data and extract meaningful information from them, thus to get a deep apprehension of the biological process.Methods: ReliefF is one efficient filter feature ranking algorithm, and it runs fast and can scale to very high dimensional datasets.It calculates the weights of the features, and sorts the features in descending order according to the weights.Since the high dimensional bioinformatics data usually contain noises and non-related information, they may affect the distances among the samples, thus influence the weights of the features calculated by ReliefF.Here we proposed a ReliefF recursive feature elimination algorithm (ReliefF-RFE), which eliminates the features that have negative weights or the feature having the smallest weight if the weights of all the features are positive in each iteration.This process will continue until the feature set is empty.The final feature rank is got according to the eliminating order.Results: One metabolomics data set about liver disease (containing 60 samples in 3 classes, and 1459 features) and two public data sets (leukemia and lymphoma, available at http://www.gems-system.org/) were used to show the validation of ReliefF-RFE.For each data set, l0-fold cross validation was run 10 times.By sequential forward searching the feature rank, the accuracy rates of the KNN classification were got.The result shows that ReliefFRFE outperforms ReliefF.Conclusions: We proposed a ReliefF-RFE feature ranking algorithm which could rank the features more accurately than ReliefF according to the feature information related to the problem .